This model is a replication of the model described in the AlexNet publication.

Differences:

• not training with the relighting data-augmentation;
• initializing non-zero biases to 0.1 instead of 1 (found necessary for training, as initialization to 1 gave flat loss).

The bundled model is the iteration 360,000 snapshot. The best validation performance during training was iteration 358,000 with validation accuracy 57.258% and loss 1.83948. This model obtains a top-1 accuracy 57.1% and a top-5 accuracy 80.2% on the validation set, using just the center crop. (Using the average of 10 crops, (4 + 1 center) * 2 mirror, should obtain a bit higher accuracy.)

This model was trained by Evan Shelhamer @shelhamer

This model is released for unrestricted use.